Contributing Multimodal Attention Patterns

[tomasruizt]

Dear @alan-cooney @neelnanda-io ! :)

I forked the repo and extended it to visualize multimodal attention patterns in vision-language-models. I would love to contribute the component to the main CircuitsVis, because it could foster more interpretability research into multimodal models.

• Here is an html example link of the new component
• I created a blog post showing how to use the component to find new research hypothesis
• This is the link to the Github fork. Atm, the component sits on top of the AttentionPatterns component, but it would probably be safer to contribute it as a separate react component.
• I'm also very happy about feedback for the component and/or for the ideas in the blog post.

Below is a screenshot of the new component. Let me know what you think :)